Existing vision-language models (VLMs) have demonstrated impressive performance in reasoning-based segmentation. However, current benchmarks are primarily constructed from high-quality images captured under idealized conditions. This raises a critical question: when visual cues are severely degraded by adverse weather conditions such as rain, snow, or fog, can VLMs sustain reliable reasoning segmentation capabilities? In response to this challenge, we introduce WeatherReasonSeg, a benchmark designed to evaluate VLM performance in reasoning-based segmentation under adverse weather conditions. It consists of two complementary components. First, we construct a controllable reasoning dataset by applying synthetic weather with varying severity levels to existing segmentation datasets, enabling fine-grained robustness analysis. Second, to capture real-world complexity, we curate a real-world adverse-weather reasoning segmentation dataset with semantically consistent queries generated via mask-guided LLM prompting. We further broaden the evaluation scope across five reasoning dimensions, including functionality, application scenarios, structural attributes, interactions, and requirement matching. Extensive experiments across diverse VLMs reveal two key findings: (1) VLM performance degrades monotonically with increasing weather severity, and (2) different weather types induce distinct vulnerability patterns. We hope WeatherReasonSeg will serve as a foundation for advancing robust, weather-aware reasoning.
Figures & tables
Figure 1: Adverse weather undermines VLM reasoning segmentation. Left: Under clean weather conditions, VLMs can correctly localize the target object and generate complete segmentation masks. However, when exposed to adverse weather, the model fails to accurately ground the reasoning query, resulting in incomplete segmentation masks and degraded pixel-level alignment and semantic reasoning performance. Right: WeatherReasonSeg addresses this limitation by introducing real-world degraded data with five structured reasoning dimensions—Function, Application, Structure, Relationship, and Requirement—to enable systematic evaluation of weather-robust reasoning.
Figure 2: Overview of WeatherReasonSeg. The framework consists of two complementary components. Top: A controllable synthetic weather generation process can generate weather type degradations (rain, snow, haze) of varying severity to construct synthetic image-query pairs for robustness evaluation. Bottom: A real-world adverse-weather reasoning dataset constructed via mask-guided large language model prompting, followed by human–model collaborative verification to ensure semantic alignment and reasoning validity.
Figure 3: Overview of WeatherReasonSeg. Left: Query word cloud. Right: Distribution of query–image pairs across weather conditions.
Figure 4: A schematic comparison of VLM reasoning segmentation results under different weather degradation levels. The VLM-based segmentation method used is Seg-Zero. The figure illustrates how segmentation performance progressively changes as weather severity increases..
Figure 5: Radar chart comparison of reasoning segmentation performance across five reasoning dimensions under different weather conditions (Fog, Rain, Snow, and Night). Results are reported using gIoU and cIoU for three representative VLM-based methods: Seg-Zero, LISA, and Seg-R1. The figure shows that visually grounded dimensions such as Structure and Function remain relatively stable, while context-dependent dimensions including Application and Requirement exhibit larger performance degradation under adverse weather conditions.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Weather
Light
Moderate
Heavy
Fog
A=0.8,β=0.32
A=0.8,β=1.2
A=0.8,β=2.0
Rain
s=200,α=0.9
s=700,α=0.85
s=1000,α=0.8
Snow
ρ=0.50,σ=3.0,γ=0.80
ρ=0.85,σ=4.0,γ=0.70
ρ=0.90,σ=4.5,γ=0.55
Appendix
Table 5: Parameter settings for different adverse weather severities.
Figure 6: Prompt construction for reasoning query generation. The prompt consists of a system instruction defining the reasoning task and a user message containing the class label, RGB image, and corresponding binary mask to guide the generation of five reasoning oriented questions.
Figure 7: The figure outlines the verification rules used to ensure that generated questions satisfy object existence certainty, single target consistency, noun-based answers, reasoning diversity, and image grounded answer ability.
Figure 8: Given the same input image and target object, we construct queries from five semantic perspectives—Function, Application, Structure, Relationship, and Requirement. The reasoning segmentation method adopts Seg-Zero. The right panel illustrates the reasoning process and segmentation predictions of the VLM model, where green check marks denote correct predictions and the red cross indicates a failure case. This example highlights how different reasoning dimensions influence the model’s ability to correctly localize the target object.
Figure 9: Given the same input image and target object, we test the reasoning and segmentation prediction results of the VLM model from five queries, where a green check indicates correct prediction and a red cross indicates prediction failure. This example highlights how different reasoning dimensions affect the model’s ability to correctly locate the target object.
Figure 10: Qualitative results of reasoning segmentation under adverse weather conditions. For each example, we show the query, the original image, and the segmentation predictions produced by Seg-Zero, Seg-R1, and LISA, together with the ground truth (GT). The results highlight the different reasoning segmentation performance of VLM based methods in environments such as rain, snow, and nighttime.
Reasoning segmentation enables vision-language models (VLMs) to translate mission-relevant language requests into pixel-level visual grounding, offering a natural perception interface for embodied agents. However, existing benchmarks largely focus on generic visual scenes and overlook the domain and resource constraints encountered in flood-response platforms. We present FloodReasonBench, a benchmark for VLM reasoning segmentation for embodied flood response at the edge. At its core, FloodReasonBench introduces FloodResponseSeg, a flood-specific reasoning-segmentation dataset constructed from real-world scenes and response-relevant targets. Beyond task accuracy, the benchmark characterizes reasoning-segmentation pipelines under lightweight visual encoding, hierarchical split inference, and compressed intermediate representations. We observe strong partition-dependent accuracy variation in the generic pre-adaptation setting, while the flood-adapted target-workload design space exhibits a substantially more compact accuracy range across partitions. Evaluation on an NVIDIA Jetson AGX Xavier further exposes the tradeoffs among reasoning-segmentation accuracy, edge-side latency, energy, and communication footprint, enabling quality-constrained selection of edge operating points. Together, these results provide a task- and system-level characterization of reasoning segmentation for resource-constrained embodied flood response at the edge.
Rajat Bhattacharjya, Yoomee Jung, Minwoo Kim +4
University of California, Irvine · Kookmin University, South Korea
This paper describes our approach for the 8th UG2+ Workshop (CVPR 2026) Track~2, which targets semantic segmentation of outdoor scenes degraded by five weather conditions: blur, darkness, snow, haze, and glare. A central challenge we observe is a severe generalization gap -- models that perform well on the validation set often collapse on the test set. For instance, SegFormer-B5 drops 16.1 mIoU points from validation to test, suggesting that model capacity alone is insufficient for robustness. We investigate whether a carefully designed training recipe, rather than architectural complexity, can address this gap. Starting from a pre-trained SegMAN-S backbone, we systematically study the effects of domain-adaptive fine-tuning, multi-source data mixing, scene-balanced sampling, and synthetic degradation augmentation. Our final system achieves 59.9% mIoU on the official test set while maintaining a validation-test gap of only 6.5 points -- less than half that of larger models. We analyze negative results from architectural modifications, loss function variants, and model scaling to provide practical insights for weather-robust segmentation under limited data.
Vision-language models (VLMs) achieve strong performance on standard, high-quality datasets, but we still do not fully understand how they perform under real-world image distortions. We present VLM-RobustBench, a benchmark spanning 49 augmentation types across noise, blur, weather, digital, and geometric perturbations, evaluated under graded severities (low/mid/high) and binary transforms, yielding 133 corrupted settings. We evaluate VLMs from four families (Qwen, InternVL, Molmo, Gemma) on two complementary benchmarks: MMBench (visually grounded) and MMMU-Pro (reasoning-oriented). Our results reveal that visual severity is a weak predictor of difficulty: low-severity spatial perturbations often degrade performance more than visually severe photometric corruptions. In particular, low-severity glass_blur reduces MMBench accuracy by about 8 pp on average across models, while the largest drops arise from resampling and geometric distortions (e.g., upsample, elastic_transform), reaching up to 34 pp. Overall, our findings suggest current VLMs are semantically strong but spatially fragile, motivating the definition of novel robustness evaluation protocols and training regimes that emphasize resampling and geometric invariances.